The Combination of Sentinel-2 MSI and Sentinel-1SAR to Estimate and Map Soil Organic Carbon Content in Farmland——A Case Study of the Typical Area of Loess hilly Area in the Huangshui Watershed of Qinghai Province

Objective The aim was to explore the potential and methods of using Sentinel-2 satellite Multi Spectral Instrument (MSI), Sentinel-1 satellite Synthetic Aperture Radar (SAR) data, and their combination to effectively estimate soil organic carbon (SOC) content in farmland, which can provide data support and services for regional SOC mapping and precision agriculture management.

Method Taking the Lanlongkou area of Huangzhong District in Xining and the Zhuozhatan area of Huzhu County in the Huangshui River Basin of Qinghai Province as a case study, soil samples were collected from farmland before spring sowing in 2022 and 2023, as well as Sentinel-2 MSI and Sentinel-1 SAR data were collected at the similar period. Partial Least Squares (PLSR), Random Forest (RF), and Extreme Gradient Boosting tree (XGBoost) models were used as estimation methods to establish a statistical model between SOC content and image spectra, and to explore the potential of combining Sentinel-2 MSI and Sentinel-1 SAR to estimate SOC content and map it.

Result Sentinel-2 was able to effectively estimate the SOC content in two regions, with a coefficient of determination（ R_V^2 ）of over 0.78, Root mean square error of validation set (RMSEv) of 1.09-1.18, and Residual Predictive Deviation (RPD) of 2.46-2.60, achieving good estimation accuracy. The R_V^2 , RMSEv, and RPD of the blocked area were 0.93, 0.84, and 4.47, respectively. The R_V^2 of the Zhuozhatan area was 0.94, RMSEv was 0.96, and RPD was 4.53, indicating that combining Sentinel-2 MSI and Sentinel-1SAR data to estimate SOC content can achieve better accuracy, and the results were better than using a single sensor. The accuracy of all three model methods met the requirements for good estimation, but the XGBoost model has the highest accuracy, better than RF and PLSR. Based on XGBoost, the SOC content in the study area was mapped, and the results showed that low SOC content mainly occurred in the valley plain area and its surrounding urban and rural residential construction land, while high SOC content areas were mainly distributed in hilly and slope cultivated areas.

Conclusion The combination of Sentinel-2 and Sentinel-1 data is an effective data source for estimating SOC content in farmland and mapping it. The research results can provide data and decision support for the fine management of farmland soil in small areas and the development of smart agriculture.